For the past month I've been doing something very unnatural to me: throwing the first one away. Going back and rewriting some aspect of a working program just to improve one property of the code, here eliminating all derived data structures and re-deriving everything all the time in a performant way.
The problem: implementing text editor operations as lines might wrap or scroll.
e.g. clicking with the mouse to reposition the cursor, or pressing the down arrow (which might cause a scroll)
to_loc: (x, y) -> loc
Identify the location at pixel coordinates (x,y) on screen.
Returns nil if (x,y) is not on screen.
to_coord: loc -> x, y
Identify the top-left coordinate on screen of location loc.
Returns nil if loc is not on screen.
down: loc, dy -> loc
Find the location at the start of a screen line dy pixels down from loc.
Returns nil if dy is taller than the screen.
Returns bottom of file if we hit it.
up: loc, dy -> loc
Find the location at the start of a screen line dy pixels up from loc.
Returns nil if dy is taller than the screen.
Returns top of file if we hit it.
hor: loc, x -> loc
Find the location at x=x0 on the same screen line as loc.
I think they might be applicable to any pixel-based editors that use proportional fonts. They seem independent of the data structure used by the editor. I use an array of lines, and so locations are defined as (line_index, pos) tuples, where pos counts in utf-8 code-points.
There's a "hard part" to programming beyond the reach of tools or methods. Sometimes a problem needs the right "algebraic" abstraction, designed around an invariant and preserving it across any composition of operations.
Not all programs get this hard.
It's useful to notice when some part calls out for doing this hard, focused work.
The code in the screenshot is a function to convert a mouse click (mx, my) into the location (line_index, pos) of the character at that point on the screen.
The problem is much of this function is boilerplate shared with several other places, such as the code to draw text on screen, compute the height of a wrapped line, etc. The boilerplate makes it difficult to see the business logic unique to this particular function, and so creates pressure to prematurely create an abstraction to "DRY things out". Highlighting the shape of the boilerplate in pink helps the eye to focus on the unique business logic in the protrusions, and so counters the pressure to hide the boilerplate before I've figured out the best way to do so.
(As an aside, this is an example of what I think of as "programmer-configured highlighting". We've gotten used to our editors deciding what to highlight for us, and we just pick the colors. One little tool I use everyday is the ability to highlight specific identifiers which flips this around: I pick a word, and the editor picks a color at random to highlight it with. And the highlight persists across sessions. The color of the State variable in the screenshot was selected in this manner.)
I've been slowly reading "The Nature of Order" by Christopher Alexander and slowly thinking about how to make my editor for text and line drawings more timeless. (And mostly sleeping a lot, if I'm honest.) Today the combination of the two led me to draw this shape for the line-wrapping algorithm.
Until now I've been developing the editor the "usual" way, which for me consists of needing some computation, figuring out the most convenient place to perform the computation, then squirreling away the result somewhere it's available when needed. In an effort to get myself out of the rut of the inevitable problems of indirection and cache invalidation that result, I've been trying to replace all my ad hoc data structures with on-demand computation based on the base state of the program. And what I've been ending up with is umpteen variations of this pictured algorithm, just with different code stuck on to the protrusions.
There may be an abstraction that comes out of all this, but I don't see it yet. And as CA says, a flower isn't made up of identical petals. Each one evolves uniquely as a part of the whole.
Here's the Lua code skeleton corresponding to that drawing. The ellipses correspond to protrusions in the drawing:
for line_index, line in array.each(State.lines, State.screen_top.line) do
if line.mode == 'text' then
local initpos = 1
if line_index == State.screen_top.line then
-- top screen line
initpos = State.screen_top.pos
end
for pos, char in utf8chars(line.data, initpos) do
if char:match('%s') then
if line_wrap_at_word_boundary(State) then
...
else
...
end
else
if x+w > State.right then
...
else
...
end
end
end
else -- drawing
...
end
end
This morning I'm thinking about turning my previous paper notation into a visual notation for something we typically use keyword args for. For example, the following glyph:
…might represent a function for initializing a text editor widget with the following signature:
Trying to build software for others is extremely disheartening. I can be eating my own dogfood on a daily basis for years and still newcomers hit bugs in their first 10 minutes.
I wonder if this is the major reason to huddle together on top of jenga stacks with tons of dependencies, terrified of fragmentation: You always need more testing than you think, and there's no way to compete with something that's been through that much testing.
The following program lets you scrub the mouse downward to find more and more precise approximations of π within the red optical sight in the center of the screen.
To recap, you basically have a line of cells that can be in one of two states ('alive' or 'dead') and rules that decide how a cell's state evolves based on the state of its immediate neighbors to the left and right. The images below show a snapshot of time in a row of pixels, and time advancing from the top row of pixels to the bottom.
Starting from a single live cell, of the 256 rules 16 immediately wink out (empty grids in the picture below), 16 don't change (vertical lines), 48 move the cell (24 each to the left and right), 30 grow into triangles over time (6 each to the left and right and 18 on both), 18 form Sierpinski patterns and 22 are more chaotic. Here's a detail in Lua Carousel where you can see many of these types.
However, things look different if you start from a random configuration of live and dead cells. Seemingly well-behaved rules hide subtleties, and seeming patterns vanish.
For a given rule, different random initial configurations largely look the same from a distance, which suggests random selection yields more realistic pictures for a rule.
Eye-balling the surface, I think 47/256 rules are chaotic.
Rule 30 is the famous one, but my favorites are rule 150 and 165.
(I've also been skimming Stephen Wolfram's "A New Kind of Science" as I do this. Wolfram separates "nested" from "random" patterns, but that seems to be an artifact of starting with a single live cell. "Nested" patterns (like Sierpinski triangles) are just a milder kind of chaos our visual cortex can get a grip on.)
Starting with a single live cell is 'simple' but grossly incomplete, exercising only scenarios 0, 1, 2, and 4 in the first step.
And if we truly care about simple, why not just start with all dead cells? Rules don't care.
Anyways, I see two fairly simple initial states that exercise every possible scenario in time step 1: the mirror images 10111 and 11101 when surrounded by runs of dead cells.
Now I see only the one real stable rule: rule 204. 204 is binary 11001100, each bit of which is exactly the middle bit of numbers 7-0. In other words, every scenario maps to the central square.
Even here, though, the long runs of dead cells keep the "true nature" of a rule from coming out. I think random initial conditions do that much better.
Perhaps what would be best is to keep our simple pattern in the center exercising all scenarios, and then pad it with a random initial state. We do have to remember to pad it out with 3 dead cells. Let's do that on both sides for symmetry.
Ah, here's a nice screenshot of a central portion of the ruleset, ideal density for chaos to emerge.
